AI Model Price vs Intelligence: September 2026 Frontier Comparison

What the September 2026 data shows

A comparison of fourteen frontier models published in September 2026 maps each model’s Artificial Analysis (AA) Intelligence Index against its blended price per million tokens. The highest score, 63.1, belongs to Claude Opus 5 at $10 per million tokens. The second-highest, Claude Fable 5 at 62.1, costs $20, the highest price in the set. At the opposite end, GLM-5.3-Flash scores 57.5 at $0.24, the lowest price. The pre-calculated ratio between the most and least expensive models reaches 84.4×.

Leaderboard and pricing tiers

Third and fourth place are shared by GPT-5.6 Sol and Grok 4.6, both at 60.9. Sol costs $8; Grok 4.6 costs $3. Fifth is Kimi K3 at 59.7 for $6. Sixth is GLM-5.3 at 59.5 for $2.15. Qwen3.8 Max and Qwen3.8 2.4T A95B follow at 58.1 and 57.7 respectively, both priced at $3. GLM-5.3-Flash at 57.5 for $0.24 marks the price floor. Muse Spark 1.2 scores 56.8 at $2. GPT-5.6 Terra reaches 56.6 at $4.5. Gemini 3.7 Flash scores 56 at $1.5. Grok 4.5 and DeepSeek V4 Pro 0813 close the table at 55.8 ($3) and 53.2 ($1.98).

Value outliers and pricing anomalies

The data exposes a non-linear premium at Anthropic. Claude Fable 5 trails Opus 5 by one index point yet costs double. Conversely, Grok 4.6 matches GPT-5.6 Sol’s intelligence at one-third the price, marking a strong price-performance outlier. GLM-5.3-Flash pushes the affordability frontier, delivering 57.5 intelligence at $0.24. The 84.4× spread reflects market segmentation: higher-priced models target enterprise use cases with stricter compliance and support requirements, while low-cost variants serve volume workloads.

Benchmark limits and hidden costs

The AA Intelligence Index is a laboratory benchmark. It does not capture full inference costs, context-window differences, or specialised capabilities such as coding and reasoning. The blended 3:1 input-to-output token ratio masks real-world cost variation; production workloads with heavy output generation will see different effective prices. API availability, service-level agreements, data residency, and model-specific benchmark performance for the intended use case all affect total cost of ownership.

Procurement guidance

For organisations prioritising maximum intelligence, Claude Opus 5 at $10 represents a more rational choice than Fable 5 at $20. Models such as GLM-5.3 ($2.15) and GLM-5.3-Flash ($0.24) deliver near-top-tier intelligence at a fraction of the price. Grok 4.6 at $3 offers a compelling alternative to higher-priced Western equivalents. The market now supplies sufficiently capable models in the single-digit dollar range per million tokens, enabling deployment by smaller organisations without paying a brand premium.